UCLA machine-learning researcher advancing LLM agent memory, memory calibration, and continual post-training
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-09) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Output data is not yet available: no publication counts, venue list, or yearly totals were provided, so this axis — which would measure the volume and cadence of peer-reviewed output over time — cannot be scored.
Impact—
Impact is also unmeasured: the snapshot contains no citation counts, h-index, or venue influence data, so this axis — citation-based influence on subsequent research — is left unscored.
Focus0/100
The focus score is 0, reflecting an unusually even field distribution: agent memory, LLM post-training, and LLM agents/NLP each account for 33.3% of the reviewed record. Read qualitatively, all three clusters sit within machine learning around LLM agents, so the low score indicates breadth across adjacent subfields rather than unrelated topic scatter.
Collaboration62/100
Collaboration scores 62, based on 22 distinct coauthors across the reviewed record. Twelve repeat collaborators appear, including Fanyu Zhao, Liang Dong, Jian Xu, and Guanjun Jiang. On all three cataloged papers Zhang holds a middle author position with no corresponding-author role, indicating steady team-based contribution within a sizable but not exceptionally wide network.
Activity—
Activity cannot be scored: no yearly publication counts or timestamps beyond three 2026 arXiv entries were provided. This axis would measure publishing cadence and recent output; current evidence covers a single year only.
Reproducibility55/100
Reproducibility readiness averages 55 across the three papers with reproducibility data. All three carry a medium rating: MemCalib ships a 15k-example benchmark that supports replication, ACLArena's comparisons are well specified but 8B-scale, and RPMem's gains are heavily ablated though evaluation is multiple-choice-only. The uniform medium tier keeps the mean moderate rather than high.
Research summary
Han Zhang is a machine-learning researcher at the University of California, Los Angeles whose reviewed work concentrates on memory and learning dynamics in large language model (LLM) agents. The Evolve catalog includes three 2026 arXiv preprints: MemCalib, a benchmark of 15k examples for measuring and optimizing memory use in LLM agents, paired with counterfactual-credit reinforcement learning; RPMem, which learns backbone-reusable recurrent parametric memory across sessions via hypernetwork-generated LoRA updates and recurrent consolidation; and ACLArena, which studies catastrophic forgetting in sequential agent post-training and compares on-policy distillation, self-distilled fine-tuning, and model merging. Evolve's reviews credit all three with thorough ablations and careful diagnosis, while flagging limits: an LLM-judge circularity in MemCalib, 8B-scale experiments and task-ID routing in ACLArena, and multiple-choice-only evaluation in RPMem. Bibliometrically, the record splits evenly across three adjacent subfields — LLM agent memory, LLM post-training, and LLM agents/NLP — with 22 distinct coauthors and a collaboration score of 62. Reproducibility readiness is rated medium for all three papers, averaging 55. On all three reviewed papers Zhang appears as a middle author without a corresponding-author role, so independent leadership is not yet evidenced by the available record.
Collaboration pattern
Han Zhang's collaboration pattern is team-based and moderately broad: 22 distinct coauthors with a collaboration score of 62, and a recurring core including Fanyu Zhao, Liang Dong, Jian Xu, Guanjun Jiang, Yifei Zhao, and Ruike Cao, among twelve repeat names. On all three cataloged 2026 papers Zhang is a middle author and never the corresponding author, consistent with contributing to shared projects rather than leading them so far.
Research keywords
Highlights
MemCalib: a 15k-example benchmark for LLM agent memory use
RPMem: recurrent parametric memory reusable across sessions
ACLArena: diagnosing forgetting in agent post-training
Timeline
- 2026Published RPMem, introducing cross-session recurrent parametric memory for LLM agents
- 2026Published ACLArena, comparing methods against forgetting in multi-stage agent post-training
- 2026Published MemCalib, benchmarking memory use in LLM agents with counterfactual-credit reinforcement learning
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Han Zhang is a common name; this profile is name-resolved from public bibliographic data and may merge or split identities.
- Bibliometric scores come from Evolve's snapshot of public bibliometric data; no citation counts, h-index, or full publication list were provided.
- The three reviewed papers are 2026 arXiv preprints; their peer-review status is unknown.
- Authorship on all reviewed papers is middle-position with no corresponding role, so research leadership is not yet supported by the current record.